[VC]AMD Fiji XT spotted at Zauba

[RussianSensation]

Using total system consumption to compare perf/watt?

OK, if you don't like that metric (because GPUs don't play in a vacuum without a CPU and motherboard), look up the performance advantage of 980 over 680/770 and you'll see 980 is nowhere near 2X faster on average unless you go 4K or multi-monitor where 2GB of VRAM destroys performance on 680/770. The power usage of a 980 is between 165 and 200W.

 

[f1sherman]

OK, if you don't like that metric (because GPUs don't play in a vacuum without a CPU and motherboard), look up the performance advantage of 980 over 680/770 and you'll see 980 is nowhere near 2X faster on average

Still derailing I see :thumbsdown:

LOL @ nowhere near

http://www.hardware.fr/articles/928-7/consommation-efficacite-energetique.html

 

[Erenhardt]

There is something about stacked on-die memory power savings, but its too late for me to find any useful numbers showing GDDR5 power consumption:
http://www.cse.psu.edu/~juz138/files/islped209-zhao.pdf

We propose an energy-efficient reconfigurable in-package graphics
memory design that integrates wide-interface graphics DRAMs with
GPU on a silicon interposer. We reduce the memory power consumption
by scaling down the supply voltage and frequency while
maintaining the same or higher peak bandwidth. Furthermore, we
design a reconfigurable memory interface and propose two reconfiguration
mechanisms to optimize system energy efficiency and throughput.
The proposed memory architecture can reduce memory power
consumption up to 54%, without reconfiguration. The reconfigurable
interface can improve system energy efficiency by 23% and throughput
by 30% under a power budget of 240W.∗

Actually, I found this. Not sure how accurate:

We computed the maximum power consumption
of GPU processors and memory controllers by subtracting
the DRAM power from the reported maximum power consumption
of Quadro R  FX5800 [15], resulting in 124W. The power of 4GB
DRAM is calculated as 60W, based on Hynix’s GDDR5 memory [8].

DRAMs on HD6990 eat 21,8% of 375 Watts card consumes = 81 Watts for memory chips alone!


There are some graphs that I'm not sure if I understand... Night all

 

[Actaeon]

^


This author agrees with me that we are stuck in no-man's land so to speak when it comes to 4K gaming right now, unless you go 3-4 high-end GPUs or turn some settings down.

I do agree that 4K should be a priority but that was not my takeaway from reading his article. He wanted a cheaper single mid-high end card capable of doing 4K smoothly with high levels of detail. He is right that we aren't there yet, but he did cite the 'super high end' of today's market as really the only thing capable of doing it. I think a high end single card is still a generation or two off from doing this. A mid-high end card maybe 3 generations.

I would say with the right combination of HW today, you could game @ 4K. The cheapest way to do this is a 290 or 290X Crossfire configuration for around $500-600. Which is what a newly released single high end GPU will cost anyway. And you can get that now instead of waiting for another generation or two with better performance than a single high end card.

HardOCP's 4K gaming reviews has been helpful for me when evaluating 4k setups. Averaging 50fps/minimum 30fps in demanding games @ 4k is plenty playable.

http://www.hardocp.com/article/2014...980_sli_overclocked_gpu_review/4#.VGp0j_ldXy4


Crossfire 390Xs will surely improve upon this and will be a killer 4K setup. Don't think a single 390X will be able to match the performance of 2x 980s or 290Xs though. Unless there is some break through, single cards are still a few generations out.

 

[Enigmoid]

Well....there is no need for stupid amounts of AA at 4K.

 

[III-V]

How much power does GDDR5 actually consume? Considering those chips are mostly passively cooled without an actual heatsink, I'd assume the GPU itself consumes (and displaces) most of the wattage required for a card.

Virtually nobody actually knows. It's similar to the saying that you need 8 glasses of water a day... no nutritionists have any idea where the saying came from, and yet everyone keeps saying it.

 

[f1sherman]

Die size can obviously be a factor in perf/W. Let's create an example where you have two chips, one 250mm2 and one 500mm2. For the sake of simplicity let's assume that the larger chip is exactly twice the smaller one, and also assume linear scaling. At the same clockspeed the 500mm2 chip is twice as fast as the 250mm2 chip. Halve the clockspeed of the 500mm2 chip and it's around around the same performance as the 250mm2 one, but the 500mm2 chip has far better perf/W because the lower clockspeed allows it to run at a far lower voltage, which is the most important adjustable factor in determining power consumption.

Basically, if you have twice the transistors doing twice the work at the same clock speed and voltage you'll have more or less 2x performance (assuming good scaling) and 2x the power consumption. If you have twice the transistors doing a bit less than twice the work, at a lower clockspeed, and at a lower voltage you'll have less than 2x performance, but better perf/w.

All true except:

Everything you said is a direct consequence of lowering the clock & voltage on bigger chip (500mm2).
You can do the same thing with smaller chip, and you'll end up with same perf/W.
And we are back to square one.

 

[raghu78]

All true except:

Everything you said is a direct consequence of lowering the clock & voltage on bigger chip (500mm2). You can do the same thing with smaller chip, and you'll end up with same perf/W. And we are back to square one.

The issue was he said the statement for a given performance. You can play around with perf/watt by lowering clocks on smaller chip ? But what about the perf ?

If you still want to claim that Maxwell brings 2x perf/watt of Kepler then GM210 must be 2x perf of GK110 at roughly similar TDP and a rumoured slightly larger die. If that does not happen then your 2x perf/watt claim is wrong.

 

[f1sherman]

For a given performance... you're right.

And if the performance is not the same (like GM204 vs GK204), then what?
We have to normalize for performance once again? LOL

Also I never claimed that die size can not be used to achieve better perf/W.
Nvidia did something to that nature with Titan/780.

But perf/W is performance per watt. It's not perf/W/area

And we can hypothesize about lowering A and increasing B, in order to gain this or that.
But there are only so much chips that make commercial sense, and that we end up buying, comparing their metrics.
And when accounted for some quirks (shader bound/bw bound, very low clocks)
perf/W seems to be a bloody constant within those real chip. It certainly does not scale well, if at all, with area.
http://www.techpowerup.com/reviews/ASUS/GTX_980_STRIX_OC/26.html

EDIT: I never claimed that. But it's not very far either. Certainly not worth getting excited over
And I explained why 2x is not possible for the new chip.

 

[f1sherman]

btw whatta crazy idea to measure perf/W for a given performance.

how about measuring just power consumption? Look ma, we invented new metrics, it's called the watt :biggrin:

 

[rtsurfer]

Wouldn't standardizing for performance work.

You set a specific score in a game or benchmark or something.
Standardize the CPU, other hardware, OS , etc.

And then you look at how much power, die area, architecture a company had to use to get there.

Couldn't a review site setup a specific rig, select a proper test and then test different GPUs.?

 

[HurleyBird]

All true except:

Everything you said is a direct consequence of lowering the clock & voltage on bigger chip (500mm2).
You can do the same thing with smaller chip, and you'll end up with same perf/W.
And we are back to square one.

More or less. Obviously some fixed units (and their power draw) are likely to stay the same between different die sizes. For example, a theoretical Maxwell chip coming in at the same size as GK104 would still require more or less the same power for the memory subsystem and board components. ROPs may or may not decrease, cache may or may not maintain the same ratio with other functional units, etc.

Like I explained in my third paragraph though, I don't think GM204's larger die size has much to do with its efficiency given that Nvidia is also running Maxwell at higher clocks vs. Kepler anyway. I'm just pointing out that die size can potentially affect power efficiency in response to those who think otherwise.

 

[raghu78]

But perf/W is performance per watt. It's not perf/W/area

I am not the one claiming Maxwell is 2x perf/watt of Kepler. If you say GM204 is 2x perf/watt of GK104 you are still wrong as performance is up 60% and power consumption is same or down 10%. So 1.6 / 0.9 = 1.77 perf/watt of GK104.

And I explained why 2x is not possible for the new chip.

GM204 and GK104 have same bandwidth. But inspite of same bandwidth you saw the improvement in perf due to better perf/shader and better memory bandwidth efficiency due to third gen delta color compression.

http://international.download.nvidi...nal/pdfs/GeForce_GTX_980_Whitepaper_FINAL.PDF

GM210 and GK110 are likely to have same bandwidth. In fact GK110 had 87.5% more shaders but only 50% more bandwidth compared to GK104. GM210 is likely to sport a 3072 cc config. so 50% more shaders with 50% more bandwidth. So if anything GM210 has a slight advantage as GK110 has lesser bandwidth per shader compared to GK104 while GM210 will have same bandwidth per shader as GM204. So your reasoning falls flat. its better to accept that you are wrong. I am not going to argue anymore with you as I have derailed the thread enough already.

 

[AtenRa]

btw whatta crazy idea to measure perf/W for a given performance.

how about measuring just power consumption? Look ma, we invented new metrics, it's called the watt :biggrin:

Efficiency(perf/watt) alone is meaningless without a context.

You can have a high perf/watt but actual performance may be low
You can have high perf/watt but actual wattage may be very high
You can have high perf/watt but the cost of the product may be too high
You can have high perf/watt but the die of the chip may be too big

Measuring efficiency(perf/watt) at a given performance or between two GPUs at the same segment(price range) or at a giver power consumption is the only way for efficiency to mean anything comparing chips or products.

Unless you only searching for the highest efficiency alone.

 

[Final8ty]

Deleted

 

[Final8ty]

Deleted.

 

[f1sherman]

Efficiency(perf/watt) alone is meaningless without a context.

You can have a high perf/watt but actual performance may be low
You can have high perf/watt but actual wattage may be very high
You can have high perf/watt but the cost of the product may be too high
You can have high perf/watt but the die of the chip may be too big

Measuring efficiency(perf/watt) at a given performance or between two GPUs at the same segment(price range) or at a giver power consumption is the only way for efficiency to mean anything comparing chips or products.

Unless you only searching for the highest efficiency alone.

Bold part: Not in the real world. No one builds for highest single-metric alone.

Perf/W alone being meaningless, and having to compare only same die chips. That's certainly a new one. What are we basing this on?

Because in the real world - perf/W seems to be characteristic of particular architecture.
And it turns out there is no correlation between die size and perf/W at all.
Every single commercially available chip has to strike the balance between all metrics, or at the very least it has to make commercial sense. And you can't achieve this by dropping ball in any single metric.
And even if you do (more hypothesizing, right?) we say: - OK fine, lets measure the end product and see where the sacrifices have been made.
We don't just say for particular metric "it makes no sense. I need such and such size chip to compare with"
So where are the sacrifices been made for before mentioned chip(s), because I see none.

Whats meaningless is normalizing against quantity with which your metrics has no correlation with.
And this is the equally meaningless and counterintuitive result that we get when insisting on such procedure.

 

[f1sherman]

GM204 and GK104 have same bandwidth.

[cut]

GM210 and GK110 are likely to have same bandwidth.

Small nuisance, but will GM200 have similar die advantage over GK110, as GM204 has over Gk104?
More important, is GM204 2x perf of GK104, like you seem to demand from GM210

Looks that having only same bandwidth is not enough for 2x perf growth. Shocking I know :ninja:

 

[DiogoDX]

Take a look on this nonsense "news": http://www.guru3d.com/news_story/sk_hynix_8ghz_3d_stackable_gddr5_memory.html

HBM = GDDR5

Tech sites.D:

 

[Erenhardt]

Take a look on this nonsense "news": http://www.guru3d.com/news_story/sk_hynix_8ghz_3d_stackable_gddr5_memory.html

HBM = GDDR5

Tech sites.D:

LOL

Also, the picture shows GDDR5 hynix memory:

2Gb 64Mx32 7.0Gbps H5GQ2H24BFR-R2C FBGA(170ball) 16Bank, 1.55V/1.55V Now

http://www.skhynix.com/inc/pdfDownload.jsp?path=/datasheet/Databook/Databook_Q4'2014_Graphics.pdf

 

[DiogoDX]

Hynix Q4 2014 porfolio:

http://www.skhynix.com/inc/pdfDownload.jsp?path=/datasheet/Databook/Databook_Q4'2014_Graphics.pdf

A page for HBM, one for GDDR5 and another for DDR3 availability and somehow they conclude that 8Gbps GDDR5 are HBM chips. LOL

 

[geoxile]

http://semiaccurate.com/forums/showpost.php?p=225159&postcount=111

Apparently Thevenin on SA is saying Carrizo will feature the Pirate Islands microarchitecture. Which I guess is IPv9 under AMD's new naming scheme since, IIRC Pirate Islands is IPv8 (Tonga and Iceland), Sea Islands IPv7, Southern Islands IPv6.

Considering the 1H ETA for both Carrizo and the next gen flagship GPU, perhaps they'll both be using Pirate Islands

 

[Silverforce11]

They definitely need GCN 2 to compete with Maxwell.

 

[ocre]

Performance per watt = performance per watt

you dont divide it by anything. It has nothing to do with die size, voltage, density, mhZ, or how many transistors. Adding in these other things is creating some random new metric.

I am not saying that die size, voltage, density, mhZ, transistors count will not have an effect on performance per watt, cause they all can. But that doesnt mean that you divide, add, multiply them into an equation and call it "performance per watt"

Performance per watt is like Body Mass Index. There are things that can have an effect on BMI. Loads and loads, such as: exercise, calories, hormones, age, pregnancy, etc
-but-
none of them are BMI.

None of these factors get divided or multiplied back in when your figuring BMI just as none of the factors that can change performance per watt get divided back in.

So when looking at performance per watt, its really simple. Performance divided by watts.

Not only has there been a huge mix up in figuring performance per watt in this thread, there also seems to be some confusion in what nvidia's marketing claims were.
When the gm204 launched, the claims were 2x performance per watt but they were specifically comparing the gtx 980 to the gtx 680.

But performance per watt is GPUs isnt constant and can change depending on the task. But, there are plenty of reviews that show that the 980 really really close to double the 680 in perf per watt. Like TPUs review shows 1.9x. And this is based on total system power consumption. Why is that important? We have a thread were a person finds that faster the GPU is the more his CPU is loaded. The GTX 980 churns out high frame rates which cause the CPU to work harder to keep up. The faster frame rates cause the CPU to use more watts too, it skews the results. There is no way around it. Isolating the GPU is incredibly difficult....but...
Even with these skewed measurements. There are plenty of reviews out there showing the gtx980 with nearly double the performance per watt of the 680.

Remember, this is not a static statistic. It is very dependent on the game/app. It fluctuates form one scenario to the next, wildly sometimes. But it is not some wild made up number.

 

[RussianSensation]

No one is arguing the definition of performance per watt. What we are saying is performance per watt is SKU specific. Making general statements that 1 architecture is X% better in performance/per watt is ONLY true if ALL SKUs on that new architecture are AT LEAST X% faster than their predecessor on the previous generation. It's that simple.

If you say Maxwell overall provides 2X the perf/watt over Kepler, then for this statement to be true, BOTH the GM204 and GM200 MUST provide at least 2X the perf/watt over GK204 and GK110, respectively.

Ok, now with that out of the way, please explain how NV's 2X perf/watt marketing statement holds up if GM200 is rumoured to be 40-60% faster than Titan Black? Do you get it now? They should have said Maxwell architecture is "Up to 2X the perf/watt" of Kepler. That way you cover all your bases.

It's a typical practice Apple uses where they will advertise something 2X faster and then you see 1-2 apps out of 1000 running 2X faster on their new GPU/CPU. So in reality it's not 2X faster but up to 2X faster. That's why I hate marketing and their entire disciple; it's full of BS to get the average consumer who has no time to do research or understand the terminology to keep buying newer products.